Estimating consistent rainfall design values for Norway using Bayesian inference and post-processing of posterior quantiles

The necessity of reliable rainfall statistics for the planning and design of infrastructure is growing as climate change leads to more frequent heavy rainfall events. These rainfall statistics are often displayed as Intensity-Duration-Frequency (IDF) curves, which depict the rainfall intensity (return level) that can be expected at a certain location over a given duration and with a certain frequency or recurrence interval (return period).

The former Norwegian computational approach for IDF curves, in use until recently, was established more than 45 years ago and has never been updated. We show that this outdated method of fitting a Gumbel distribution to the heaviest precipitation events fails to reflect the return values associated with long return periods. Instead, we introduce a method based on Bayesian inference for fitting a Generalized Extreme Value (GEV) distribution to observed annual maximum precipitation for several target durations. Generally, the proposed method performs well. However, the resulting IDF curves for some stations may be inconsistent across durations and return periods due to the estimation being carried out independently across durations. To avoid this, we subject the IDF curves to a post-processing where a quantile selection algorithm searches for consistent return levels within the posterior quantiles of the Bayesian inference. The post-processing produces consistent estimates that are at least as accurate as the unadjusted, inconsistent estimates.

Unlike before, the resulting new IDF curves feature uncertainty intervals that are derived directly from the Bayesian approach. The curves are presented to the users on the homepage of the Norwegian Centre for Climate Services (https://klimaservicesenter.no).